Table 4. Comparison of feature importance measures for respective binomial classification models.
Feature | Chinese–Japanese | Chinese–English | Japanese–English | |||
---|---|---|---|---|---|---|
RFI (GBDT) | PFI (MLP) | RFI (GBDT) | PFI (MLP) | RFI (GBDT) | PFI (MLP) | |
Speechiness | 4.8 | 1.05 | 46.6 | 1.67 | 39.6 | 1.22 |
Loudness | 14.1 | 1.46 | 3.0 | 1.19 | 4.7 | 1.04 |
Instrumentalness | 26.0 | 1.22 | 2.8 | 1.08 | 8.5 | 1.105^ |
Acousticness | 31.1 | 1.16 | 34.1 | 1.35 | 2.4 | 1.04 |
Energy | 32.9 | 1.07 | 4.7 | 1.09 | 16.0 | 1.114^ |
Mode name | 0.0 | 1.00 | 0.0 | 1.01 | 0.2 | 1.00 |
Duration | 9.3 | 1.08 | 4.3 | 1.07 | 12.9 | 1.02 |
Danceability | 1.2 | 1.01 | 1.6 | 1.06 | 10.4 | 1.110^ |
Valence | 0.7 | 1.06 | 1.2 | 1.02 | 3.3 | 1.003 |
Tempo | 1.4 | 1.01 | 1.5 | 1.02 | 1.9 | 1.00 |
Liveness | 0.0 | 1.00 | 0.2 | 1.00 | 0.1 | 1.00 |
Note:
While scales between RFI and PFI are not equivalent, both measure model-specific feature importance relative to other features: the higher the score, the larger the importance within the model. Features with highest importance are in bold. PFIs were reported with two decimal places, but we used three decimal places for PFIs denoted by ‘^’. This was to identify the 2nd most important feature for the PDP.